基于Gabor卷积和Transformer的课堂表情识别研究  

Research on Classroom Expression Recognition Based on Gabor Convolutional and Transformer

作　　者：曲大鹏[1] 杨天奇 郭伟嘉 田芳茵 QU Da-peng;YANG Tian-qi;GUO Wei-jia;TIAN Fang-yin(Faculty of Information,Liaoning University,Shenyang 110036,China;Graduate School,Liaoning University,Shenyang 110036,China)

机构地区：[1]辽宁大学信息学部,辽宁沈阳110036 [2]辽宁大学研究生院,辽宁沈阳110036

出　　处：《辽宁大学学报（自然科学版）》2024年第3期208-219,共12页Journal of Liaoning University：Natural Sciences Edition

基　　金：辽宁省研究生教育教学改革研究资助项目(LNYJG2022012);辽宁省教育科学‘十四五’规划项目(JG21DB242);辽宁省教育科学‘十三五’规划项目(JG20DB199);辽宁大学研究生教学改革重点项目(YJG202301055)。

摘　　要：为了针对复杂环境变化无法精准识别学生表情问题,本文设计了一个基于Gabor卷积和Transformer的表情识别模型Gabor-Vision-Transformer(GVT).将Gabor卷积和Transformer的思想相结合,设计了一个特征提取块GVT-Block.首先通过Gabor卷积提取富含丰富纹理和边缘信息的面部局部特征,再通过Transformer提取全局数据之间的长距离信息,从而更好地学习面部关键特征,显著提高模型的分类效果.GVT模型在RAF-DB和FER2013Plus数据集上的准确率分别为88.56%和87.38%,并与多个模型进行对比实验和分析,验证了本模型效果的优越性.This article mainly studies student expression recognition in complex environments.In response to the problem of complex environmental changes that cannot accurately recognize person′s expressions,a facial expression recognition model GVT(Gabor-Vision-Transformer)based on Gabor convolution and Transformer is designed.A feature extraction block GVT block was designed by combining Gabor convolution with the idea of Transformer.By using Gabor convolution to extract local facial features rich in texture and edge information,and then using Transformer to extract long-distance information between global data,we can better learn facial key features and significantly improve the classification performance of the model.The accuracy of GVT on the RAF-DB and FER2013Plus datasets is 88.56%and 87.38%,respectively.Comparative experiments and analysis with multiple other models have verified the superiority of this model.

关 键 词：表情识别 Gabor卷积网络 TRANSFORMER 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]